Application of LSTM Models in Predicting Particulate Matter (PM2.5) Levels for Urban Area

Abstract

Air pollution in India poses a big threat to human lives. In 2017, 77% of population of India was subjected to PM2.5 (Particulate Matter) exposure resulting in mortality of 6.7 lakh throughout the country. In this study, Long Short-Term Memory (LSTM) model, a powerful deep learning technique is applied for PM2.5 prediction. Three variants of LSTM model, LSTM for regression, LSTM for regression using window and LSTM for regression with time steps are developed to predict PM2.5 concentration in India. The metrics used to evaluate the performance of the predictive models are root mean square error (RMSE) and coefficient of determination (R2). The models are applied to continuous ambient air quality data collected from 14 stations in India, for the period from May 01, 2019 to April 30, 2020 at an interval of every 15 minutes. The optimal results are obtained from the models with the tuned parameters of 64 epochs and batch size of 32. All the three variants of LSTM model performed equally well in predicting PM2.5 concentration. The experimental results revealed that the value of R2 is maintained at 0.9 consistently for all the variants of LSTM model. The low values of RMSE and high values of R2 proved the reliability of the model. Thus, the proposed model gives awareness about the air pollution level in India and alerts the society to take precautionary steps to save their lives. Further the urban planners can have idea of the pollution levels for their planning and decision making.

Multivariate air quality time series analysis via a recurrent variational deep learning model

“…Prior (6) where the decoding structure follows that of the traditional SSM, and the prior follows an AR structure with increasing order, facilitated by an RNN. The graphical structure of the generative model is shown in Fig.…”

“…Numerous discriminative techniques based on deep learning models have been proposed for AQI prediction. [1][2][3][4][5][6][7] Most works focus on short-term or one-step-ahead prediction. Additionally, they tend to predict the AQI value, which is a function of the 6 indicator concentrations.…”